Explainable Machine Learning for Characterizing Unknown Molecular Structures in Infrared Spectra

Various machine learning methods for detecting functional groups in infrared (IR) spectra have been developed to accelerate time-consuming and labor-intensive IR spectrum analysis. Nevertheless, existing machine learning methods have exhibited two fundamental limitations: (1) Existing methods do not incorporate prior knowledge of IR spectrum analysis during the training and inference processes. (2) Machine learning approaches are inherently black-box models, and therefore, the prediction results of existing methods lack explainability. To overcome these limitations, we propose the substructure-directed spectrum interpreter network (SSIN), a deep learning method for efficient and explainable functional group detection in IR spectra of unknown molecules. SSIN achieved a detection accuracy greater than 0.920 on the NIST database, which contains experimentally measured 8845 gas-phase IR spectra. Furthermore, SSIN accurately identified the absorption peaks corresponding to the existence of target functional groups. Quantitatively, human-readable IR spectrum analysis reports generated by SSIN with large language models showed 81-99% accuracy when compared with the ground-truth IR spectroscopy absorption table. Source codes and optimized models of SSIN are publicly available at https://github.com/ngs00/SSIN.